On Faithfulness and Factuality in Abstractive Summarization

Summary

A 2020 ACL paper that conducts a large-scale human evaluation of hallucinated content in neural abstractive summarisation systems on the eXtreme Summarization (XSum) task. The authors compare RNN, CNN, Transformer and pretrained-LM summarisers as well as human-written summaries, asking how frequently each hallucinates and what kinds of hallucinations they produce. They also evaluate whether automatic metrics correlate with the human faithfulness judgements.

Contribution

Two pieces. (1) A documented finding that "human annotators found substantial amounts of hallucinated content in all model generated summaries," but that pretrained models are better summarisers — not only by ROUGE but also by human-judged faithfulness and factuality. (2) A methodological finding: textual-entailment-based measures correlate with faithfulness better than standard ROUGE-style metrics, pointing toward better automatic evaluation, training and decoding criteria.

Method

Large-scale human evaluation on XSum across multiple model families (RNN, CNN, Transformer, pretrained), plus correlation analysis of candidate automatic metrics against human faithfulness judgements. Annotations released publicly.

Relevance to RISE

A foundational pre-LLM reference on hallucination in text-generation that anchors the hallucination and reasoning-faithfulness themes for RISE. Catalog projects that synthesise or summarise sources at scale — storm, autosurvey, surveyx, paper-qa — inherit exactly the intrinsic/extrinsic hallucination failure modes this paper documents, and the entailment-as-evaluation finding directly informs how their output should be measured.

Critique / open questions

Conducted on XSum, an extreme-summarisation benchmark not representative of research-paper-length generation; pre-LLM era, so quantitative numbers do not directly transfer to current models. Faithfulness is defined relative to a single source document, not to multi-source synthesis as in RISE pipelines.

Key quotes

"These models are highly prone to hallucinate content that is unfaithful to the input document."

"Textual entailment measures better correlate with faithfulness than standard metrics, potentially leading the way to automatic evaluation metrics as well as training and decoding criteria."